AI and machine learning interview questions

Target encoding replaces a category with a statistic of y, such as the mean label.

One-hot becomes huge, so people use target encoding, hashing, embeddings, or frequency encoding.

Interactions help linear and logistic models capture AND-style effects they cannot learn from main effects alone.

If missingness itself predicts the target, mean imputation hides that signal.

SMOTE builds synthetic minority points by interpolating between a minority row and its neighbors.

Class weights leave the original rows intact and only change the loss.

It plots precision against recall as you move the threshold, focusing on the rare class.

Threshold tuning is often the first and cheapest lever when probabilities are usable.

ColumnTransformer applies different transformers to different column subsets in one step.

FeatureUnion runs transformers on the same data and concatenates their outputs.

Pass the whole Pipeline to GridSearchCV or RandomizedSearchCV.

Backpropagation applies the chain rule to compute the gradient of the loss with respect to every weight.

In deep sigmoid or tanh stacks, gradients shrink as they go backward, so early layers barely learn.

An epoch is one full pass over the training set.

Elastic net mixes L1 and L2.

For vanilla SGD they match: both add a term that pulls weights toward zero.

Stopping before the optimizer fully minimizes train loss limits effective capacity.

Use stratified k-fold for classification so each fold keeps the class mix.

GroupKFold keeps all rows from the same group, such as a user or a hospital, on one side of the split.

ShuffleSplit draws repeated random train and val cuts, which is flexible when you want a fixed train size.